THE DETERMINANTS OF HOUSEHOLD SAVING IN CHINA:A DYNAMIC PANEL ANALYSIS OF PROVINCIAL DATA
Charles Yuji HoriokaJunmin Wan
Working Paper 12723http://www.nber.org/papers/w12723
NATIONAL BUREAU OF ECONOMIC RESEARCH1050 Massachusetts Avenue
Cambridge, MA 02138December 2006
We are grateful to Barry Bosworth, Hideki Hayashi, Takako Fujiwara-Greve, Teh-Ming Huo, InsangHwang, Shinsuke Ikeda, Junichiro Ishida, Miki Kohara, Justin Yifu Lin, Ronald I. McKinnon, KazuoOgawa, Hugh Patrick, Masaya Sakuragawa, Shizuka Sekita, Katsuya Takii, Midori Wakabayashi,Xiaoping Wang, Calla Weimer, Tongsheng Xu, Zhihao Yu, Yaohui Zhao, and especially ChristopherCarroll, Galina Hale, Aart Kraay, Louis Kuijs, Colin McKenzie, Masao Ogaki, and Etsuro Shioji, andparticipants of the Seoul Conference on ?China and Emerging Asia: Reorganizing the Global Economy,?the Seattle Conference of the Asia-Pacific Economic Association, the Summer Institute of the NationalBureau of Economic Research, the Annual Pacific Basin Conference of the Federal Reserve Bankof San Francisco, and the fall meeting of the Japanese Economic Association, seminars at the OsakaSchool of International Public Policy (OSIPP ) of Osaka University, the Asian Public Policy Programof Hitotsubashi University, the China Center for Economic Research of Peking University, the Schoolof Economics of Jiangxi University of Finance and Economics, the Faculty of Economics of KeioUniversity, the Faculty of Economics of Fukuoka University, and the Brookings Institution for theirvaluable comments, and Horioka is grateful to the Ministry of Education, Culture, Sports, Scienceand Technology of the Japanese Government for Grant-in-Aid for Scientific Research number 18330068,which supported this research. The views expressed herein are those of the author(s) and do not necessarilyreflect the views of the National Bureau of Economic Research.
The Determinants of Household Saving in China: A Dynamic Panel Analysis of ProvincialDataCharles Yuji Horioka and Junmin WanNBER Working Paper No. 12723December 2006JEL No. D12,D91,E21,J10
ABSTRACT
In this paper, we conduct a dynamic panel analysis of the determinants of the household saving ratein China using a life cycle model and panel data on Chinese provinces for the 1995-2004 period fromChina?s household survey. We find that China?s household saving rate has been high and rising andthat the main determinants of variations over time and over space therein are the lagged saving rate,the income growth rate, (in many cases) the real interest rate, and (in some cases) the inflation rate. However, we find that the variables relating to the age structure of the population have the expectedimpact on the household saving rate in only one of the four samples. These results provide mixedsupport for the life cycle hypothesis as well as the permanent income hypothesis, are consistent withthe existence of inertia or persistence, and imply that China?s household saving rate will remain highfor some time to come.
Charles Yuji HoriokaInstitute of Social and Economic ResearchOsaka University6-1 Mihogaoka, Ibaraki-shiOsaka-fu 567-0047 JAPANand [email protected]
Junmin WanOsaka School of International Public PolicyOsaka University1-31, Machikaneyama-choToyonaka, Osaka 560-0043 [email protected]
2
1. Introduction
China has attracted increasing attention because it is the world’s most populous
nation and because it has maintained phenomenal rates of economic growth in recent
years. For example, the Asian Development Bank now projects that China will attain a
growth rate in excess of 9% in 2006 for the fifth consecutive year, thereby serving as the
engine of growth in the Asian-Pacific region (Nihon Keizai Shimbun, evening edition of
April 6, 2006, page 1).
Moreover, another reason for being interested in China is that China introduced
a so-called “one-child policy” in 1979 as a way of controlling population growth. This
is an interesting natural experiment that makes fertility largely exogenous and enables
us to assess the impact of the age structure of the population on the household saving
rate without worrying about endogeneity issues. Moreover, because the one-child
policy was applied more leniently to ethnic minorities, the policy also led to substantial
variations among provinces in the age structures of their populations, and this will
enable us to more sharply estimate the impact of the age structure of the population on
the household saving rate.
Yet another noteworthy aspect of China’s economy is its high saving rate.
China has had by far the highest overall saving rate in the world since at least 2000, and
her saving rate has increased even further since 2000—to nearly 50% of GDP. Gross
capital formation (investment) is also high in China, but because saving exceeds
investment, China has been running a net saving surplus, which translates into a current
account surplus, and that surplus has been growing sharply--from 1.9% of GDP in 2000
to 3.6% in 2004 and a remarkable 7.2% in 2005--even though China is investing at a
staggering rate of 43-46% of GDP and even though China is still relatively poor. This
3
has made China one of the world’s largest capital exporters and has exacerbated trade
frictions with the United States and other countries. Moreover, China’s net saving
surplus shows no signs of abating (The Economist, September 24-30, 2005, edition,
page 13 of “A Survey of the World Economy”).１ Thus, it is important to understand
the determinants of, and future trends in, China’s saving rate, and the obvious
candidates are the rapid rates of economic growth alluded to earlier and the age
structure of the population, which has shown tremendous variation over time as well as
over space.
In this paper, we conduct a dynamic panel analysis of the determinants of the
household saving rate in China using a life cycle model and panel data on Chinese
provinces for the 1995-2004 period from China’s household survey.
At least two previous studies have conducted similar analyses. Kraay (2000)
uses panel data on Chinese provinces from China’s household survey to analyze the
determinants of the saving rates of rural and urban households during the 1978-83 and
1984-89 periods and finds that, in the case of rural households, future income growth
has a negative and significant impact on their saving rates, that the share of food in total
consumption has a negative and significant impact on their saving rates, presumably
because households closer to the subsistence level have less ability to save, and that
neither the dependency ratio (proxied by the ratio of population to employment) nor
future income uncertainty has a significant impact on their saving rates. However,
Kraay (2000) finds that virtually none of the explanatory variables has a significant
impact on the saving rates of urban households. Modigliani and Cao (2004) conduct a
regression analysis of the determinants of the household saving rate using times series
data for the 1953-2000 period and find that the long-term growth rate, the reciprocal of
4
the dependency ratio (proxied by the ratio of the employed population to the number of
minors), the deviation of growth from the long-term growth rate, and inflation all have
positive and significant impacts on the household saving rate. Thus, the two studies
obtain somewhat conflicting results. Kraay (2000) finds that the dependency ratio
does not have a significant impact on the household saving rate, whereas Modigliani
and Cao (2004) find that it does. Moreover, Kraay (2000) finds that future income
growth has a negative and significant impact on the household saving rate, whereas
Modigliani and Cao (2004) find that the long-term growth rate and the deviation of
growth from the long-term growth rate have a positive and significant impact on the
household saving rate.
The current study improves upon these earlier studies in a number of respects:
(1) the data are much newer, (2) the dependent variable (the household saving rate) is
defined more carefully and includes household investments in real assets, (3) the
dependency ratio is defined more carefully and the young dependency ratio and the old
dependency ratio are entered separately, (4) we include variables not included by
previous authors such as the lagged saving rate and the interest rate, (5) we obtain
results for the sample of urban households, the sample of rural households, the sample
of all households, and a pooled sample of urban and rural households (unlike Kraay
(2000), who obtains results only for urban and rural households, and Modigliani and
Cao (2004), who obtain results only for all households), and (6) we use superior
estimation techniques.
This paper is organized as follows: In section 2, we present data on household
saving rates and related variables; in section 3, we present our theoretical model; in
section 4, we discuss the estimation model and data sources; in section 5, we discuss the
5
estimation method; in section 6, we present the estimation results; and section 7 is a
concluding section.
To preview our main findings, we find that China’s household saving rate has
been high and rising and that the main determinants of variations over time and over
space therein are the lagged saving rate, the income growth rate, (in many cases) the
real interest rate, and (in some cases) the inflation rate. However, we find that the
variables relating to the age structure of the population have the expected impact on the
household saving rate in only one of the four samples. These results provide mixed
support for the life cycle hypothesis as well as the permanent income hypothesis, are
consistent with the existence of inertia or persistence, and imply that China’s household
saving rate will remain high for some time to come.
2. Data on Saving Rates and Other Related Variables
In this section, we present data on household saving rates and other related
variables.
First, Figure 1 shows data on trends over time in the age structure of the
population during the 1949-2004 period, and as can be seen from this figure, there have
been pronounced trends over time in both the young dependency ratio (the ratio of the
population aged 0-14 to the population aged 15-59) and the old dependency ratio (the
ratio of the population aged 60 or older to the population aged 15-59). The former
increased from 0.57 in 1950 to 0.77 in 1964 before starting to decline, falling to 0.28 by
2004 (due in large part to the “one-child policy” and other population control measures),
while the latter increased more or less steadily from 0.13 in 1950 to 0.18 in 2004.
Finally, the total dependency ratio (the ratio of the population aged 0-14 or 60 or older
6
to the population aged 15-59) showed more or less the same trends over time as the
young dependency ratio, increasing from 0.70 in 1950 to 0.89 in 1964 before starting to
decline, falling to 0.46 by 2004 (also due in large part to the “one-child policy” and
other population control measures). The life cycle hypothesis predicts that the age
structure of the population will have a significant impact on the saving rate and in
particular that the dependency ratios will have a negative impact on the saving rate, and
if we compare trends over time in the national saving rate with trends over time in the
dependency ratios, the upward trend in the saving rate that has been observed since the
1960s coincides with a downward trend in the young and total dependency ratios during
the same period, suggesting that the latter may be a cause of the former.
Looking next at the age structure of China’s population in international
comparison, China’s young dependency ratio was higher than the worldwide level in
1975 (0.74 vs. 0.67) but fell at an unprecedented rate due to the one-child policy and
other population control measures. As a result, it was far less than the worldwide level
by 2005 (0.32 vs. 0.46).２
By contrast, the old dependency ratio was somewhat lower than the worldwide
level in 1975 (0.13 vs. 0.16) but has gradually increased due to the steady increases in
life expectancy and was just under the worldwide level by 2005 (0.16 vs. 0.17).
However, because trends over time in the young dependency ratio have been
more pronounced than trends over time in the old dependency ratio, trends in the total
dependency ratio mirror trends in the youth dependency ratio: it was just over the
worldwide level in 1975 (0.87 vs. 0.83) but declined sharply thereafter, falling to far
less than the worldwide level by 2005 (0.48 vs. 0.63).
The fact that the young and total dependency ratios were formerly relatively
7
high by international standards can explain why China’s saving rate was formerly
relatively low by international standards, and the fact that the young and total
dependency ratios are now relatively low by international standards can explain why
China’s saving rate is now relatively high by international standards.
Figure 2 shows data on trends over time in the saving rates of urban, rural, and
all households for the 1995-2004 period from China’s household survey, and as can be
seen from this figure, the saving rates of the three categories of households are roughly
comparable not only with respect to their levels but also with respect to trends over time
therein. Looking first at the level of the saving rate, the saving rates of urban, rural,
and all households fluctuated in the 17.33% to 22.98%, 15.78% to 29.77%, and 16.29%
to 25.17% ranges, respectively, and averaged 20.3%, 24.7%, and 22.4%, respectively,
during the 1995-2004 period. The fact that the saving rate of rural households is
considerably higher than that of urban households even though their income levels are
so much lower is surprising, but it could be due to the greater income volatility of rural
households, the vast majority of whom are farmers, as a result of which they save more
for precautionary purposes, or to the fact that differences in income levels largely reflect
differences in price levels, as a result of which the purchasing power of the incomes of
urban and rural households is not nearly as different as their nominal incomes.
Turning to trends over time in the saving rates of urban, rural, and all
households, the saving rate of urban households showed an upward trend throughout the
1995-2004 period, while the saving rates of rural and all households showed upward
trends until 1999 before leveling off slightly. The upward trends in the saving rates of
all three categories of households coincide with the downward trends in the young and
total dependency ratios, and thus it is possible that the latter are one of the causes of the
8
former. Thus, the evidence presented thus far suggests that the age structure of
China’s population can explain not only the high level of China’s household saving rate
but also the upward trend therein.
Table 1 shows data on the average saving rates of urban, rural, and all
households during the 1995-2004 period by province, and as can be seen from this table,
there has been enormous variation among provinces in their saving rates, with the
saving rate of urban households ranging from 10.7% to 25.7%, that of rural households
ranging from 10.0% to 43.7%, and that of all households ranging from 13.5% to 35.1%.
Finally, Table 2 shows data on the age structure of urban, rural, and all
households by province during the 1995-2004 period, and as can be seen from this table,
there has been enormous variation among provinces in the age structure of their
populations as well. For example, the young dependency ratio ranged from 0.17 to
0.39 for urban households, from 0.18 to 0.52 for rural households, and from 0.18 to 0.48
for all households, the old dependency ratio ranged from 0.07 to 0.18 for urban
households, from 0.07 to 0.16 for rural households, and from 0.07 to 0.18 for all
households, and the total dependency ratio ranged from 0.29 to 0.48 for urban
households, from 0.34 to 0.66 for rural households, and from 0.31 to 0.56 for all
households. We will conduct a regression analysis in sections 5 through 7 to see if
variations in the household saving rate correlate with variations in the age structure of
the population.
3. Theoretical Model
From the above sections, we know that the Chinese economy can be
characterized as showing a high saving rate, rapid economic growth, and a dramatic
9
change in the age structure of her population. In this section, we present a
theoretical model that provides the justification for our estimation model. In
particular, we present a two-period overlapping generations model based on the life
cycle hypothesis and use this model to show the impact of income growth and the
age structure of the population on the saving rate.
Suppose that there is an endowment economy. The time horizon is from 0
to T. The economy starts from time 0. At time 0, 0n (=1) persons with an
endowment or human capital of 0y (=1) enter the economy, and at time 1, they
become old and die. This person earns no wages at time 1 but she has to consume
at both time 0 and time 1, so she has to save at time 0 to finance her consumption at
time 1. The person has no bequest motive. The variables n and y denote
population growth and income growth, respectively, and are positive.
Consequently, at period t, tn persons newly enter the economy with endowment
tt yn and live together with 1−tn old persons. It is also assumed that the tn new
persons cannot lend to or borrow from the 1−tn old persons. All persons
homogenously have a log linear utility function with a subtractive stock of habits
and/or durability. For simplicity, the rate of time preference and the interest rate
are assumed to be zero, but we have confirmed that the qualitative findings remain
unchanged even when a positive rate of time preference and a positive interest rate
are assumed. Thus, the new generation solves the following individual
maximization problem at time t:
ttt
ttt
yccts
ccc
=+
−+
+
+
1
1
..
)log()log(max γ
where tc , 1+tc denotes consumption at time t and time t+1, respectively. γ is a
10
parameter that represents the degree of durability or habit formation, with a negative
value of gamma indicating the existence of durability, a positive value indicating the
existence of habit formation, and a zero value indicating the absence of both
durability and habit formation. We also assume that γ>-0.5 in order to insure that
0>tc , 01 >+tc . Given these assumptions, we obtain the following solutions for
consumption at time t and time t+1:
,22 γ+
=t
tyc
(1)
.22)21(
1 γγ
++
=+
t
tyc (2)
Next, we derive the aggregate saving rate in this economy at time t (t>0) as
follows:
.112221
,22)21(
22
11
⎟⎟⎠
⎞⎜⎜⎝
⎛−
++
=
⎟⎟⎠
⎞⎜⎜⎝
⎛+
++
+−
=
−−
ny
yn
ynynynSR tt
tt
tttt
γγ
γγ
γ (3)
The impacts of y, n and γ on the aggregate saving rate are as follows:
,0)22(
212 >
++
=∂∂
nyySR
γγ (4)
,0)22(
212 >
++
=∂∂
ynnSR
γγ (5)
.10
,1
,10
,)1(212
<<
==
>>
+−
=∂∂
nyfor
nyfor
nyfor
nynySRγγ
6)
These results imply that income growth and an increase in the working-age
11
population will raise the saving rate and that the strengthening of habits or the
weakening of durability will raise the saving rate in a booming economy but lower it
in a declining economy. These implications from the theoretical model suggest that
China’s high and rising household saving rate may be due to rapid income growth,
the decline in the dependency ratio, and/or the uniqueness of utility parameters.
Compared with the previous literature, the impact of γ on individual saving
(individual saving = 1+tc , 0)22(/ 21 >+=∂∂ −+ γγtc ) is the same as that of, for
example, Lahiri and Puhakka (1998). However, the impact of γ on aggregate
saving as expressed in equation (6) cannot be found in previous studies.
4. Estimation Model and Data Sources.
In this section, we discuss the estimation model and data sources we use in our
empirical analysis.
The dependent variable we use in our analysis is SR = the household saving
rate, defined as the ratio of household saving to household disposable income (net
household income in the case of rural households) and where household saving is
calculated as household disposable (or net) income minus household consumption.
Following Loayza, et al. (2000) and Schrooten and Stephan (2005), we
estimate a reduced-form linear equation rather than adhering to a particular, narrow
structural model, but the theoretical literature offers guidance regarding what variables
should be included as explanatory variables. Since the life cycle hypothesis predicts
that the household saving rate will be a function of the growth rate of per capita income
and the age structure of the population (see section 3 above and also Modigliani (1970)
and Deaton (1992), Chapter 2), we include the following explanatory variables:
12
(1) CHY = the income growth rate, defined as the real rate of growth of per
capita household disposable income (net household income in the case of rural
households)
(2) YOUNG = the young dependency rate, defined as the ratio of the
population aged 0-14 to the population aged 15-64３
(3) OLD = the old dependency rate, defined as the ratio of the population aged
65 or older to the population aged 15-64
(4) DEP = the total dependency rate, defined as the ratio of the population aged
0-14 or 65 or older to the population aged 15-64
In addition, we include the following explanatory variables:
(5) SR(-1) = the one-year lag of the saving rate
(6) RINT = the real interest rate, defined as NINT – INFL, where NINT = the
nominal interest rate on one-year bank deposits and INFL = the rate of change of the
consumer price index
(7) INFL = the rate of change of the consumer price index
(8) RURAL = a dummy variable that equals 1 in the case of rural households
and zero otherwise (included only when the pooled sample of urban and rural
households is used)
(9) A constant term
The lagged saving rate is included to test for the presence of inertia or
persistence. The real interest rate is included to test for the impact of financial
variables, and we would expect its coefficient to be positive if the substitution effect
13
more than offsets the income effect. The inflation rate is included as a proxy for price
uncertainty and/or macroeconomic stability more generally (as done by Loayza, et al.
(2000) and Schrooten and Stephan (2005)), a rural dummy is included to see if there are
any systematic differences between urban and rural households in trends over time in
the household saving rate, and a constant term, which corresponds to the coefficient of
the time trend in the regressions in differences (see section 6 below), is included in
some variants
Finally, the real growth rate of per capita gross provincial product is used as an
instrument in the level equation, as discussed below.
The data we use in our analysis are panel data for 1995-2004 on Chinese
provinces. All variables are available for urban, rural, and all households with the
exception of the nominal interest rate, which is available only for the country as a whole,
and the real growth rate of per capita gross provincial product, which is available only
for each province as a whole. Thus, we are able to obtain separate results for urban,
rural, and all households and for a pooled sample of urban and rural households.
All data from China’s household survey and national accounts data are taken
from the China Statistics Yearbook, all demographic data are taken from the China
Population Statistics Yearbook, and data on nominal interest rates are taken from the
International Monetary Fund’s International Financial Statistics.
Data were available for all 31 provinces for the ten-year period from 1995 to
2004 with the following exceptions: data were not available for Chongqing Province
during the 1995-96 period because this province did not become independent of
Sichuan Province until 1997, and data on the CPI and/or on household income and
consumption were not available for Tibet Province during the 1995-98 period. These
14
missing values caused the number of observations to decline from 310 to 304.
Moreover, one year’s worth of data were lost because the income growth rate was used
as an explanatory variable. This reduced the number of observations further from 304
to 273 and means that the sample period for most provinces was nine years (1996-2004).
Finally, because the lagged real growth rate of per capita gross provincial product was
used as an instrument, yet another observation for Chongqing Province (that for 1998)
had to be dropped, causing the final number of observations to be 272.
Descriptive statistics on the variables used in our analysis for the final sample
of 272 observations are shown in Table 3.
5. Estimation Method
In this section, we briefly describe our estimation method. Following Loayza,
et al. (2000) and Schrooten and Stephan (2005), we use a
generalized-method-of-moments (GMM) estimator applied to dynamic models using
panel data. We use this estimator for at least three reasons: (1) Inertia is likely to be
present in annual data, and it seemed desirable to use a dynamic specification to allow
for it. (2) Some of the explanatory variables (such as RINT and CHY) are likely to be
jointly determined with the saving rate, and it seemed desirable to control for the
potential joint endogeneity of the explanatory variables. (3) There is the possibility of
unobserved province-specific effects correlated with the regressors, and it seemed
desirable to control for such effects.
Following Loayza, et al. (2000) and Schrooten and Stephan (2005), we use the
alternative “system GMM estimator” proposed by Arellano and Bover (1995) and
Blundell and Bond (1998), which reduces the potential biases and imprecision
15
associated with the usual difference estimator by combining, in a system, the regression
in differences with the regression in levels.
As Windmeijer (2005) notes, the estimated asymptotic standard errors of the
efficient two-step GMM estimator will be severely downward biased in small samples,
and thus we correct the standard errors for this bias using the method proposed by
Windmeijer (2005).４
Following Loayza, et al. (2000) and Schrooten and Stephan (2005), the
demographic variables (YOUNG, OLD, and DEP) are the only explanatory variables
that we treated as being strictly exogenous and included as instruments in the level
equation as well as the first-difference equation. All other explanatory variables were
regarded as being weakly exogenous, and lagged values thereof were included as
“internal instruments,” with Bond’s (2002) method being used to select instruments.５
Finally, the one-period lag of the real growth rate of per capita gross provincial product
was used as an instrument only in the level equation. All of the instruments we use
passed all of the commonly used tests: the Hansen test, the AR(1) test, and the AR(2)
test. Tables 4-6 show the results of these tests and also show which instruments were
used in each equation.
6. Estimation Results
In this section, we present our estimation results concerning the determinants
of the household saving rate. The estimation results for urban, rural, and all
households and for a pooled sample of urban and rural households are shown in Tables
4, 5, 6, and 7, respectively.
Looking first at the coefficient of SR(-1) (the lagged saving rate), this
16
coefficient is always positive and highly significant, indicating strong inertia or
persistence. This coefficient ranges from 0.476 to 0.628, implying a long-run effect
that is 1.91 to 2.69 times the short-run effect, in the sample of urban households; from
0.476 to 0.844, implying a long-run effect that is 1.91 to 6.41 times the short-run effect,
in the sample of rural households; from 0.622 to 0.721, implying a long-run effect that
is 2.65 to 3.58 times the short-run effect, in the sample of all households; and from
0.604 to 0.710, implying a long-run effect that is 2.54 to 3.45 times the short-run effect,
in the pooled sample of urban and rural households.
Looking next at the coefficient of CHY (the income growth rate), it is always
positive and highly significant (which is consistent with the life cycle hypothesis),
ranging from 0.192 to 0.270 in the sample of urban households, from 0.331 to 0.536 in
the sample of rural households, from 0.201 to 0.240 in the sample of all households, and
from 0.337 to 0.396 in the pooled sample of urban and rural households. These figures
imply that a one percentage point increase in the income growth rate causes a 0.192 to
0.536 percentage point increase in the household saving rate. Moreover, the long-run
impact of the income growth rate is 1.91 to 6.41 times these figures.
Looking next at the coefficient of RINT (the real interest rate), it is
insignificant and usually positive in the sample of urban households, positive and
significant in two out of four cases in the sample of rural households, and positive and
significant in all cases in the sample of all households and the pooled sample of urban
and rural households. Thus, the real interest rate has a significant positive impact on
the household saving rate for every sample except for the sample of urban households,
which suggests that the interest elasticity of saving is positive and is consistent with the
permanent income hypothesis.
17
Looking next at the impact of the demographic variables (YOUNG, OLD, and
DEP), their coefficients are never significant in the sample of urban households and the
sample of rural households, the coefficients of YOUNG and DEP are positive and
sometimes significant but the coefficient of OLD is insignificant in the sample of all
households, and the coefficients of YOUNG and DEP are negative and at least
marginally significant but the coefficient of OLD is insignificant in the pooled sample
of urban and rural households. Thus, the only case in which the coefficients of the
demographic variables are significant with the expected sign is in the case of the
coefficients of YOUNG and DEP in the pooled sample of urban and rural households,６
and the coefficients of YOUNG and DEP are sometimes significant with the wrong sign
in the sample of all households. The reasons for the mixed performance of the
demographic variables is a topic for future research.７
Looking next at the coefficient of INFL (the inflation rate), it is insignificant in
the sample of urban households, the sample of all households, and the pooled sample of
urban and rural households, but it is negative and significant in three out of four cases in
the sample of rural households. These results suggest that the impact of inflation is not
always significant but that it is sometimes negative and significant.
Looking next at the coefficient of the RURAL dummy in the pooled sample of
urban and rural households, it is positive and significant in all four cases, which
suggests that the saving rate of rural households have a more pronounced upward trend
than that of urban households after controlling for other factors.
Looking finally at the constant term, which represents the coefficient of a time
trend, it is positive in all cases and is significant in five out of eight cases, which
suggests that there is an upward trend in China’s household saving rate.
18
We also tried adding year dummies and the level of per capita household
disposable income as additional explanatory variables, but we dropped them from the
final specification because their coefficients were not statistically significant.
Lastly, we compare our results to those of previous studies. Our finding that
income growth has a positive and significant impact on the household saving rate is at
variance with Kraay’s (2000) finding that (future) income growth has a negative and
significant impact on the saving rate of rural households and does not have a significant
impact on the saving rate of urban households but is consistent with Modigliani and
Cao’s (2004) finding that (long run) income growth has a positive and significant
impact on the household saving rate. In order to shed light on why our results differ
from those of Kraay (2000), we tried estimating all of our equations using two-stage
least squares, the same estimation method used by Kraay (2000), and found that the
results are substantially different. For example, the coefficients of the variables
relating to the age structure of the population, which had previously been insignificant,
are now significant, whereas the coefficient of income growth, which had previously
been positive and significant, becomes totally insignificant (which is consistent with
Kraay’s (2000) results for urban households) when two-stage least squares are used.
These findings suggest that the differences between our results and those of Kraay
(2000) are due largely to differences in estimation method and underscore the
importance of using dynamic panel techniques when using panel data.
7. Conclusion
In this paper, we conducted a dynamic panel analysis of the determinants of the
household saving rate in China using a life cycle model and panel data on Chinese
19
provinces for the 1995-2004 period from China’s household survey. To summarize our
main findings, we found that China’s household saving rate has been high and rising
and that the main determinants of variations over time and over space therein are the
lagged saving rate, the income growth rate, (in many cases) the real interest rate, and (in
some cases) the inflation rate. However, we found that the variables relating to the age
structure of the population have the expected impact on the household saving rate in
only one of the four samples. These results provide mixed support for the life cycle
hypothesis (with the positive and significant coefficient of income growth supporting
the life cycle hypothesis and the mixed performance of the demographic variables being
unfavorable to the life cycle hypothesis), provide some support for the permanent
income hypothesis (with the positive and significant coefficient of the interest rate
supporting this hypothesis), and are also consistent with the existence of inertia or
persistence.
Turning to the implications of our findings, our finding that inertia or
persistence are strong implies that there will not be a dramatic decline in China’s
household saving rate, and our finding that the income growth rate has a positive impact
on the household saving rate implies that China’s household saving rate will remain
high as long as the growth rate remains high. However, if the growth rate tapers off,
we can explain a gradual decline in the household saving rate.
Thus, it seems likely that China’s household saving rate will remain high in the
short to medium run, and to the extent that this causes China’s current account surplus
to remain high, this may cause continued frictions with the United States and China’s
other trading partners. In the long run, however, China’s household saving rate can be
expected to taper off assuming the growth rate tapers off, and thus, in the long run,
20
China may well suffer from current account deficits rather than current account
surpluses.
Turning finally to directions for further research, there are a number of factors
that we were not able to consider in this analysis due to data limitations, such as
borrowing constraints, precautionary saving, bequest motives, the distribution of
income, and old-age pensions, health insurance, and other social insurance programs,
and we hope to be able to incorporate these factors in our future research.
21
References
Alessie, Rob, and Annamaria Lusardi. (1997) “Consumption, Saving and Habit Formation.” Economics Letters, 55:1 (August 15), 103-108. Arellano, Manuel, and Stephen Bond. (1991) “Some Tests of Specification for Panel Data: Monte Carlo Evidence and an Application to Employment Equations.” Review of Economic Studies, 58:2 (April), 277-297. Arellano, Manuel, and Olympia Bover. (1995) “Another Look at the Instrumental Variable Estimation of Error-Components Models.” Journal of Econometrics, 68:1 (July), 29-51. Banister, Judith. (1984) “An Analysis of Recent Data on the Population of China.” Population and Development Review, 10:2 (June), 241-271. Banister, Judith. (1987) China’s Changing Population. Stanford, California: Stanford University Press. Blundell, Richard, and Stephen Bond. (1998) "Initial Conditions and Moment Restrictions in Dynamic Panel Data Models." Journal of Econometrics, 87:1 (August), 115-143. Bond, Stephen. (2002) “Dynamic Panel Data Models: A Guide to Micro Data Methods and Practice.” CEMMAP Working Paper CWP09/02, Department of Economics, Institute for Fiscal Studies, London. Chamon, Marcos, and Eswar Prasad. (2006) “Determinants of Household Saving in China.” Mimeo., International Monetary Fund, Washington, D. C. Deaton, Angus. (1992) Understanding Consumption. Oxford: Clarendon Press. “The Frugal Giant.” The Economist, September 24-30, 2005, issue, pp. 12-14 of “A Survey of the World Economy.” Kraay, Aart. (2000) “Household Saving in China.” World Bank Economic Review, 14:3
22
(September), 545-570. Kuijs, Louis. (2005) “Investment and Saving in China.” World Bank Policy Research Paper Series No. 3633 (June). Available at SSRN: http://ssrn.com/abstract=756985. Kuijs, Louis. (2006) "How Will China's Saving-Investment Balance Evolve?" World Bank Policy Research Working Paper No. 3958 (July 1). Available at SSRN: http://ssrn.com/abstract=923265. Loayza, Norman; Klaus Schmidt-Hebbel; and Luis Serven. (2000) “What Drives Private Saving across the World?” Review of Economics and Statistics, 82:2 (May), 165-181. Modigliani, Franco. (1970) “The Life Cycle Hypothesis of Saving and Intercountry Differences in the Saving Ratio.” In Induction, Growth and Trade: Essays in Honour of Sir Roy Harrod, edited by W. A. Eltis, M. FG. Scott, and J. N. Wolfe, pp. 197-225. Oxford: Clarendon Press. Modigliani, Franco, and Shi Larry Cao. (2004) “The Chinese Saving Puzzle and the Life-Cycle Hypothesis.” Journal of Economic Literature, 42:1 (March), 145-170. Roodman, D. (2005) “ xtabond2: Stata module to extend xtabond dynamic panel data estimator.” Center for Global Development, Washington, D.C. http://econpapers.repec.org/software/bocbocode/s435901.htm Roodman, D. (2007), “How to do xtabond2: An introduction to ‘difference’ and ‘system’ GMM in Stata,” Working Paper 103, Center for Global Development, Washington, D.C. http://www.cgdev.org/content/publications/detail/11619 Schrooten, Mechthild, and Sabine Stephan. (2005) “Private Savings and Transition: Dynamic Panel Data Evidence from Accession Countries.” Economics of Transition, 13:2, 287-309. Windmeijer, Frank. (2005) “A Finite Sample Correction for the Variance of Linear Efficient Two-Step GMM Estimators.” Journal of Econometrics, 126:1 (May), 25-51.
23
Data Sources
Central Intelligence Agency, CIA World Factbook. http://www.indexmundi. com/g/g.aspx?c=ch&v=30 Department of Population, Social, Science and Technology Statistics, National Bureau of Statistics of China, ed., China Population Statistics Yearbook, 1991-2005 editions. Beijing: China Statistics Press. International Monetary Fund，International Financial Statistics, 1995-2005. National Bureau of Statistics of China., ed., China Statistical Yearbook, 1988-2005 editions. Beijing: China Statistics Press. United Nations (2002), World Population Prospects: The 2002 Revision. New York: United Nations.
24
Figure 1 : Age Structure of the Popu lation in Ch ina, 1949-2004
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1949
1954
1959
1964
1969
1974
1979
1984
1989
1994
1999
2004
Sources: China Population Statistics Yearbook, 1988-2005 editions, and World PopulationProspects: The 2002 Revision (United Nations)
Ratio of thepopulationaged 0-14 tothe populationaged 15-59
Ratio of thepopulationaged 60 orolder to thepopulationaged 15-59
Ratio of thepopulationaged 0-14 or60 or older tothe populationaged 15-59
25
Figure 2: Household Saving Rate in China, 1995-2004
0.00
5.00
10.00
15.00
20.00
25.00
30.00
35.00
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
Source: Authors' calculations based on China Statistics Yearbook, 1996-2005 editions
Source: Authors' calculations based on China Population Statistics Yearbook, 1996-2005 editions.
Table 2: Age Structure of the Population by Province (Averagesfor the 1995-2004 Period)
Notes: The young dependency ratio is defined as the ratio of the population aged 0-14 to thepopulation aged 15-64; The old dependency ratio is defined as the ratio of the population aged 65 orolder to the population aged 15-64; The total dependency ratio is defined as the ratio of thepopulation aged 0-14 or 65 or older to the population aged 15-64.
Source: Authors' calculations based on China Statistics Yearbook, 1996-2005 editions, ChinaPopulation Statistics Yearbook, 1996-2005 editions, and International Financial Statistics, 1995-2005 editions.
Number of observations 272 272 272 272Number of groups 31 31 31 31Hansen test of overidentification (p-value) 0.737 0.540 0.978 0.522Test for 1st-order serial correlation (p-value) 0.000 0.000 0.000 0.001Test for 2st-order serial correlation (p-value) 0.121 0.131 0.124 0.126Transformation used
Instruments only for first difference equationGMM(SR(-1), CHY,RINT, INFL, (2 .)collapse)
GMM(SR(-1), CHY,RINT, INFL, (2 .)collapse)
GMM(SR(-1), CHY,RINT, INFL, (3, 3))
GMM(SR(-1), CHY, RINT,INFL, (2 .) collapse)
Instruments for both first difference and levelequations
YOUNG, OLD DEP YOUNG, OLD DEP
Instruments only for level equation
Notes: Standard errors are in parentheses; *, **, *** denote significant at the 10%, 5% and 1% levels, respectively.
Table 4: The Determinants of the Household Saving Rate in China (Urban Households)
Dependent variable = SR
first differences
CHGDP(-1)
30
SR(-1) 0.774 0.844 0.481 0.476
(0.068)*** (0.042)*** (0.168)*** (0.169)***
CHY 0.495 0.536 0.332 0.331
(0.107)***
(0.149)***
(0.181)*
(0.155)**
RINT 0.591 0.593 0.069 -0.022
(0.163)***
(0.185)*** (0.369) (0.492)
YOUNG 0.030 -0.088
(0.033) (0.075)
OLD 0.200 -0.001
(0.157) (0.243)
DEP 0.025 -0.078
(0.024) (0.069)
INFL -0.356 -0.338 -0.773 -0.843
(0.180)* (0.232) (0.284)
**(0.370)
**
Constant 0.164 0.173
(0.052)*** (0.063)**
Number of observations 272 272 272 272Number of groups 31 31 31 31Hansen test of overidentification (p-value) 0.596 0.538 0.636 0.613Test for 1st-order serial correlation (p-value) 0.000 0.000 0.013 0.014Test for 2st-order serial correlation (p-value) 0.680 0.802 0.496 0.506Transformation used
Instruments only for first difference equationGMM(SR(-1), CHY, RINT,INFL, (2 .) collapse)
GMM(SR(-1), CHY, RINT,INFL, (2 .) collapse)
GMM(SR(-1), CHY, RINT,INFL, (2 .) collapse)
GMM(SR(-1), CHY, RINT,INFL, (2 .) collapse)
Instruments for both first difference and levelequations
YOUNG, OLD DEP YOUNG, OLD DEP
Instruments only for level equation
Notes: Standard errors are in parentheses; *, **, *** denote significant at the 10%, 5% and 1% levels, respectively.
first differences
CHGDP(-1)
Table 5: The Determinants of the Household Saving Rate in China (Rural Households)
Dependent variable = SR
31
SR(-1) 0.721 0.711 0.658 0.622
(0.028)***
(0.034)***
(0.097)***
(0.119)***
CHY 0.201 0.204 0.211 0.240
(0.073)*** (0.097)** (0.097)** (0.109)**
RINT 0.513 0.622 0.435 0.500
(0.157)***
(0.154)***
(0.144)***
(0.270)*
YOUNG 0.058 0.053
(0.026)** (0.034)
OLD 0.070 0.077
(0.068) (0.078)
DEP 0.058 0.034
(0.025)** (0.025)
INFL 0.113 0.145 0.048 -0.003
(0.119) (0.115) (0.120) (0.193)
Constant 0.018 0.034(0.033) (0.039)
Number of observations 272 272 272 272Number of groups 31 31 31 31Hansen test of overidentification (p-value) 1.000 0.187 1.000 0.209Test for 1st-order serial correlation (p-value) 0.000 0.000 0.000 0.001Test for 2st-order serial correlation (p-value) 0.262 0.251 0.277 0.226Transformation used
Instruments only for first difference equationGMM(SR(-1), CHY, RINT,INFL, (2 4))
GMM(SR(-1), CHY, RINT,INFL, (2, 5) collapse)
GMM(SR(-1), CHY,RINT, INFL, (2, 4))
GMM(SR(-1), CHY, RINT,INFL, (2, 5) collapse)
Instruments for both first difference and levelequations
YOUNG, OLD DEP YOUNG, OLD DEP
Instruments only for level equation
Notes: Standard errors are in parentheses; *, **, *** denote significant at the 10%, 5% and 1% levels, respectively.
first differences
CHGDP(-1)
Table 6: The Determinants of the Household Saving Rate in China (All Households)
Note: Young Dependency Ratio is defined as the ratio of the population aged 0-14 to the population aged 15-59; OldDependency Ratio is defined as the ratio of the population aged 60 or older to the population aged 15-59; TotalDependency Ratio is defined as the ratio of the population aged 0-14 or 60 or older to the population aged 15-59.
Sources: China Population Statistics Yearbook, 1988-2005 editions; Banister (1987); World Population Prospects: The2002 Revision (United Nations); and U.S. CIA Factbook.
Data Appendix for Figure 1 and Other Related Variables
Source: Authors' calculations based on China Statistics Yearbook, 1996-2005 editions.
Data Appendix for Figure 2
35
Endnotes １ See Kuijs (2005, 2006) for data on the overall level and sectoral composition of
saving and investment and on the saving-investment balance in China. ２ The demographic data in this paragraph and the two following paragraphs are based
on United Nations data and hence do not coincide precisely with the earlier data. ３ It would have been preferable to use the population aged 15-59 since the retirement
age in China (for males) is 60, but we could not do so due to data limitations. ４ All calculations were done using Stata, version 10. We used Roodman’s (2005,
2007) ”xtabond2” program in Stata to correct the standard errors. ５ The “collapse” command in Stata was used to select instruments. ６ The fact that the coefficients of YOUNG and DEP are negative and significant only in
the pooled sample is presumably due to the fact that the one-child policy has been
enforced much more strictly in urban areas than in rural areas, as a result of which there
is much greater variation in YOUNG and DEP in the pooled sample than in the other
samples. ７ Chamon and Prasad (2006) analyze micro data from the same household survey we
use and find that saving increases with age and is highest for the elderly. This can help
explain why OLD and DEP do not have the expected impacts on the household saving
